arxiv:1902.03381v2 [physics.pop-ph] 15 feb 2019cultural history. the painting portrays a night sky...
TRANSCRIPT
Is The Starry Night Turbulent?
James R. Beattie1 and Neco Kriel2
1Research School of Astronomy and Astrophysics, Australian National University, Canberra,Australia
2Science and Engineering Faculty, Queensland University of Technology, Brisbane, Australia
Vincent van Gogh’s painting, The Starry Night, is an iconic piece of art andcultural history. The painting portrays a night sky full of stars, with eddies(spirals) both large and small. Kolmogorov (1941)’s description of subsonic,incompressible turbulence gives a model for turbulence that involves eddiesinteracting on many length scales, and so the question has been asked: is TheStarry Night turbulent? To answer this question, we calculate the azimuthallyaveraged power spectrum of a square region (1165 × 1165 pixels) of night sky inThe Starry Night. We find a power spectrum, P(k), where k is the wavevector,that shares the same features as supersonic turbulence. It has a power-lawP(k) ∝ k−2.1±0.3 in the scaling range, 34 ≤ k ≤ 80. We identify a driving scale,kD = 3, dissipation scale, kν = 220 and a bottleneck. This leads us to believe thatvan Gogh’s depiction of the starry night closely resembles the turbulence found
in real molecular clouds, the birthplace of stars in the Universe.
1. INTRODUCTION
Recently, parallels have been drawn between theKolmogorov (1941)’s turbulence (i.e. subsonic,incompressible, high Reynolds number flow) and theswirls and vortices depicted by Vincent Van Gogh, inparticular, van Gogh’s The Starry Night (Aragon et al.,2008). The Starry Night is of particular interest to theastronomer who knows that stars are born inside ofturbulent molecular gas clouds (Scalo, 1998; Ferriere,2001; Mac Low & Klessen, 2004; Kainulainen et al.,2009; Arzoumanian et al., 2011; Federrath & Klessen,2012; Federrath, 2013; Andre et al., 2014; Federrath,2016; Hacar et al., 2018; Mocz & Burkhart, 2018). If,by coincidence, van Gogh had constructed a turbulentstarry night this would be a pleasing coincidence.Previously, Aragon et al. (2008) established that therewere scaling relations present in the luminosity values ofThe Starry Night. However, a detailed power spectrumanalysis is still lacking, which is essential to establish ifThe Starry Night is indeed turbulent.
The aim of this study is to perform this analysis bycalculating the azimuthally-averaged power spectrum,P(k), where k is the wavevector, of the night sky inThe Starry Night. This will be essential for establishingwhether or not there exists something that resembles anenergy cascade, P(k) ∝ kα, with a Kolmogorov power-law scaling, α = −5/3, a driving, kD and a dissipationscale, kν , (Kolmogorov, 1941). All of which are keyfeatures that define turbulent flows.
This article is organised into the following sections:In §2 we follow Kolmogorov (1941)’s derivation of thescaling laws of turbulent flows, and indicate which
1contact: [email protected]: [email protected]
quantities we will need to measure in order understandif The Starry Night indeed depicts turbulent behaviourin the night sky. Next, in §3 we discuss the pre-processing methods that we use on The Starry Night. In§4 we perform a two-dimensional (2D) power spectrumanalysis to determine if the eddy structures are isotropic(an assumption of Kolmogorov 1941). Following we usethe 2D power spectrum to construct the azimuthally-averaged power spectrum, providing us with a meansof directly testing if The Starry Night is turbulentin nature (i.e. if a scaling range, driving scale anddissipation scale exist). Finally, in §5 we summariseour findings.
2. TURBULENCE THEORY
In this section we follow Kolmogorov (1941)’s (K41)original arguments to clarify what it means to beturbulent. We first consider eddies (vortices) on lengthscale `, with rotation velocity v`, which have a turnovertime t` ∼ `/v`. We denote the kinetic energy of theeddy at length scale ` with ε`, which means the flowrate of the eddy is approximately,
ε` =dε`dt∼ ε`t`∼(v2`2
)(v``
)∼ v3`
`. (1)
If dissipation is negligible, i.e. for high Reynoldsnumber (Re ∼ v/ν, where ν is the dissipation), thenε` is conserved and the kinetic energy of the turbulenteddies flows from small to large length scales. Thisdefines the energy cascade (or scaling range, which ismore general),
L > ` > `ν , (2)
where L is the largest scale in the turbulence, and
Is The Starry Night Turbulent?
arX
iv:1
902.
0338
1v2
[ph
ysic
s.po
p-ph
] 1
5 Fe
b 20
19
2 J. R. Beattie & N. Kriel
FIGURE 1. Vincent van Gogh’s The Starry Night, accessed from WallpapersWide.com (2018). We see eddies paintedthrough the starry night sky that resemble the structures comparable to what we see in turbulent flows.
`ν is the scale which dissipation effects are no longernegligible. This leads to the scaling relation v ∼ (ε`)1/3,of the velocity through the energy cascade. We will besearching for such a scale in van Gogh’s Starry Night,however, we will not be searching in `, but rather in k,the wavevector space. To construct the above argumentin wavevector space we must now introduce the 2nd
order structure function.The 2nd order structure function is the average of the
square difference in velocities separated by distance `.It is mathematically defined as,
SF2(`) = 〈[v(r + `)− v(r)]2⟩
=⟨(δv)2
⟩, (3)
where the operator 〈. . .〉 indicates the averagingover the ensemble of independent positions x. If theturbulence is statistically isotropic, a question that wewill explore in this study, then the averages of δv2
only depend upon the magnitude of the separationdistance, |`|. Clearly SF2(|`|) ∝
⟨v2`⟩
and therefore,from Equation 1,
SF2(|`|) ∝ (ε`)2/3. (4)
The power spectrum, which is further discussed in §4,is related to the velocity fluctuations by,
1
2
∫ ∞0
〈v′`〉 d3`′ =
∫ 0
∞Pv(k)d3k. (5)
where k is the wavevector ` = 2π/k. Usingthe previous proportionality relations and taking theFourier transform, we find,
Pv(k) ∝ SF2 `3 ∝ k−3(k−1ε2/3) ∝ ε2/3k−11/3. (6)
For isotropic turbulence we know that d3k → 4π2dk.This means that
Pv(k)d3k ∝ Pv(k)k2dk ∝ ε2/3k−5/3dk (7)
for power in the modes from k to k + dk. The k−5/3
scaling is a prediction made by the K41 model. Tobe clear, we assume that we were looking at scaleson L > ` > `ν , that the flow was incompressible(i.e. no ρ dependence), and that the flow wasstatistically isotropic. For supersonic spectrum, asimilar power-law, k−2, was predicted (modellingsupersonic turbulence as a series of shocks, which haveFourier transforms ∼ 1/k) and tested with confirmationin high resolution simulations (Burgers, 1948; Kritsuket al., 2007; Konstandin et al., 2012; Federrath, 2013;Federrath et al., 2018). The aim of this study will be toidentify the scaling range (if one exists) and calculatethe power-law P(k) ∝ kα.
Is The Starry Night Turbulent?
Is The Starry Night Turbulent? 3
FIGURE 2. The square (1165× 1165) region, separated by channel, that we selected from The Starry Night to encompassthe largest eddy structures, whilst omitting the city and the towering structure in the left of Figure 1
3. STARRY-NIGHT DATA
To analyse the vortices in Van Gogh’s Starry Nightwe first acquire a high resolution digital copy of thepainting. We access a 2560 × 1600 pixel image fromWallpapersWide.com (2018), shown in Figure 1. Thereare more structures (i.e. buildings and trees) in thepainting than just (possibly) turbulent vortices. Forthis reason, we focus on a region on the painting thatcontains the night sky. This limits us to the analysis ofthe largest square region in the night sky, containing themaximum amount of vortice information. We choose asquare region (1165 × 1165) because then constructingthe power spectrum becomes an easier task. The regionwe select is shown in Figure 2. All three of the channelscontain structural information about the nature of thevortices, as shown in Figure 2, so we perform our 2Dpower spectrum analysis on each of them. In previousstudies, averaging is performed at this stage in theprocessing across the channels, however we restrainourselves from averaging until we know that there isnot a significant difference between them. From hereinwe will refer to the square region of The Starry Nightas SN(`), where ` ∈ R2 is the coordinate vector in real-space. Also note whilst K41’s model is based upon thepower spectrum of the velocity, we can only access pixelvalues, similar to the analysis in Aragon et al. (2008),hence SN(`) is the pixel value at pixel position `.
4. POWER SPECTRUM METHODS
4.1. The 2D Fourier Transform and PowerSpectrum
To access the scaling relations discussed in §2 weconstruct the 2D power spectrum for the vortices shownin Figure 2.
First we calculate the 2D Fourier transform,
SN(k) =1
n2
∫SN(`)e−ik·`d2`, (8)
where n = 1165 is the box size of the region in TheStarry Night. The power spectrum is the square of theFourier transform, analogous to the regular definitionof power, which is the square of intensity of a signal.It is therefore given as the square of the intensities offrequencies in the 2D Fourier space, hence we calculateit by,
P(k) = SN(k)SN∗(k), (9)
where SN∗(k) is the complex conjugate of the Fourier
transform. In this study we encode the normalisationof power spectrum into the coefficient of the Fouriertransform, shown in equation 8. For this reason,equation 8 deviates slightly from the regular definitionof the 2D Fourier transform. Finally, we centre thepower spectrum by shifting the k = (0, 0) to the centreof the 2D data.
The overall goal here is to reduce the 2D powerspectrum into a one-dimensional (1D) power spectrumby performing azimuthal averaging, so that we canidentify any features that resemble turbulent dynamics.We can only justify averaging over the azimuthal anglefor particular |k| if the spectrum is isotropic, i.e.invariant under rotations about the origin. To test thiscondition, we first apply a Gaussian smoothing kernel(with standard deviation σ = 10) to the 2D powerspectra, to reduce the Gaussian noise. Once we havesmoothed the spectrum, we perform a marching squaresalgorithm on the spectrum to reveal the isobars i.e. thecurves of equal power. The smoothing allows us to justfocus on the overall shape of the isobars and gives us aqualitative means of observing if the spectra is isotropic.
4.2. 2D Power Spectrum Results
In the top three panels of Figure 3 we plot the2D power spectra for each of the channels shown infigure 2. In the bottom three panels we shown thesmoothed power spectra, with isobars identified using
Is The Starry Night Turbulent?
4 J. R. Beattie & N. Kriel
FIGURE 3. Top: the centred, 2D power spectrum, P(k) for each of the channels. Bottom: the 2D power spectrumconvolved with a Gaussian kernel to smooth the spectrum, and then shown with isobars every ≈ 0.5 log10 P(k). We see thatthe power spectrum is ansiotropic for small k, but becomes more isotropic as k � 1.
the marching squares algorithm. For large k, andsmall `, we see a relatively isotropic power spectrum.Perhaps this is because it is easiest to paint symmetric,small structures, since these can be performed in asingle brush stroke. For small k and hence large `(towards the origin of the bottom three panels in Figure3) we see the power spectrum becomes stretched inthe ky direction. This suggests that van Gogh hasa preferential orientation in the way he painted largestructures in his starry night, possibly from a particulartechnique that he used to paint it.
Figure 3 also suggests that the power spectra fromeach channel are relatively similar in structure andmagnitude, especially on small scales. This is good,because we can average over the power spectra for eachchannel to produce a single 2D power spectrum, whichin turn can be used to create a single azimuthally-averaged power spectrum for The Starry Night, whichwe do next.
4.3. The Azimuthally-Averaged Power Spec-trum
Now that we have observed that the 2D power spectrumis similar across each channel we average across them,
creating a single 2D power spectrum for the night skyin The Starry Night. We now use this single powerspectrum to construct the azimuthal average.
We construct concentric circles, each of radius |k| =k, separated by a single wavevector magnitude, fromk = 0 to 821. We average over all powers at each of thefixed k, through a θ ∈ [0, 2π] rotation in the 2D k-space,producing a single power value for each k,⟨
P(k)⟩θ
=⟨SN(k)SN
∗(k)2πk
⟩. (10)
Where the operator 〈. . .〉θ is the azimuthal-averagearound the k-space circle.
We also construct the variance,
σ2 =⟨P2(k)
⟩θ−⟨P(k)
⟩2θ, (11)
which we use to construct the 1σ uncertainties in thepower spectrum.
4.4. Azimuthally-Averaged Power SpectrumResults
In Figure 4 we plot the azimuthally-averaged powerspectrum from the 3-channel average of the night sky
Is The Starry Night Turbulent?
Is The Starry Night Turbulent? 5
FIGURE 4. Top: The azimuthally averaged power spectrum of van Gogh’s Starry Night (shown in black) with 1σuncertainties constructed from the variance of the averaging for each |k| = k (shown in blue). The power spectrum hassimilar characteristics to what one would expect from a power spectrum of a field undergoing turbulent dynamics. A drivingscale, kD, and a dissipation scale, kν , can be identified. We can also identify a region in the k-space that seems to follow apower-law, resembling a scaling range where P(k) ∼ k−2.1±0.3, indicated by the red dashed lines. Bottom: The left plot isscaled by k−5/3, to see if the scaling range becomes stationary under a K41 power-law, and the right plot is scaled by k−2,to see if the scaling range becomes stationary under Burgers’ power-law. We find a small negative gradient under the K41scaling, but complete stationarity under the Burgers (1948) scaling.
in The Starry Night. We see a peak in the spectrumat small k, which we identify as a driving scale kD = 3.Moving to larger k we see a dip, and then what mightbe a power-law between k = 34 to k = 80. Weidentify this as a scaling-range, noting that it seemsto follow a power-law under the log-log transform, asone expects from both K41 and Burgers (1948). At stilllarger k we find a dissipation scale kν = 220, where theazimuthally-averaged power spectrum quickly tapersoff. Before the dissipation scale, more prominentlyshown the two bottom plots in Figure 4, we also seea build-up of energy at approximately kν . This isknown as the “bottle-neck effect”, and is realised in realturbulence (Dobler et al., 2003).
We plot the averaged power spectra with k−5/3
and k−2 compensations in two bottom plots in Figure4. Identifying the scaling range, we fit the functionlogP(k) = α log k+β, and find that α = 2.1±0.3. Thisis why the bottom left plot (K41 scaling) does not quitebecome stationary, showing a small, negative slope,whereas in the bottom right plot (Burgers’ scaling)shows a close to zero slope. α = 2.1 ± 0.3 is is alsoconsistent with the results from Aragon et al. (2008),who finds that SF2(`) ∝ `p, where p ≈ 2 (Aragon et al.2008 uses R instead of `). This suggests that the nightsky from The Starry Night is a coincidental depictionof supersonic turbulence (P(k) ∝ k−2), a key featureof real molecular clouds in the Universe, which are thebirth places of stars (Burgers, 1948; Kritsuk et al., 2007;Konstandin et al., 2012; Federrath & Klessen, 2012;
Is The Starry Night Turbulent?
6 J. R. Beattie & N. Kriel
Federrath, 2013; Andre et al., 2014; Federrath, 2016;Hacar et al., 2018; Mocz & Burkhart, 2018).
5. SUMMARY AND KEY FINDINGS
In this study we determined whether or not the nightsky in van Gogh’s Starry Night has a power spectrumthat resembles a supersonic turbulent flow. First weselect a square region of the night sky from The StarryNight. Next, we calculate the two-dimensional (2D)power spectrum for each of the three red-green-blue(RGB) channels. We find the channels have simi-lar power spectra and average over the three spectrato create a single spectrum. We then construct theazimuthally-averaged power spectrum by averagingover concentric circles of radii |k| = k, where k ∈ Zis the wavevector. We observe a driving scale, kD,dissipation scale, kν , and scaling range that followsa power-law similar to that of supersonic turbulence.This shows that van Gogh’s The Starry Night doesexhibit some similarities to turbulence, which happensto be responsible for the real, observable, starry nightsky. We summarise the key findings below:
• The 2D power spectrum of the night sky in TheStarry Night is mostly invariant to the RGB chan-nel. It is isotropic for large k, suggesting that vanGogh’s night-sky has small, symmetrical struc-tures. For small k, corresponding to large scalesin the painting, the power spectrum is squeezed inthe ky direction. This may be due to a particulartechnique that van Gogh used to paint the largeeddies in the night-sky.
• We identify characteristic scales in the azimuthally-averaged power spectrum that may resemble adriving scale at k = 3, a dissipation scale atk = 220 and scaling range between 34 ≤ k ≤ 80.We also identify a “bottleneck-effect” at approxi-mately kν . These realisations support the case thatThe Starry Night is a depiction of a turbulent flow.
• The slope of the power spectrum in the identifiedscaling range is −2.1 ± 0.3, similar to the scalingexpected in supersonic turbulence, −2. The valueof the slope agrees with the second order structurefunction length scaling found in Aragon et al.(2008), suggesting that van Gogh’s depiction of thestarry night sky is in fact an accurate portrayal of asupersonic, turbulent medium, bursting with stars.
The code for this study can be found at the GitHubrepository:https://github.com/AstroJames/
VanGoghsStarryNight.git
ACKNOWLEDGEMENTS
J. R. B. would like to acknowledge the useful discussionshe had with Dean Muir, David Galea and, the briefdiscussion of the power spectrum results with ChristophFederrath.
REFERENCES
Andre P., Di Francesco J., Ward-Thompson D.,Inutsuka S.-I., Pudritz R. E., Pineda J. E., 2014,Protostars and Planets VI, pp 27–51
Aragon J. L., Naumis G. G., Bai M., Torres M., MainiP. K., 2008, Journal of Mathematical Imaging andVision, 30, 275
Arzoumanian D., et al., 2011, A&A, 529, 1
Burgers J., 1948, Advances in Applied Mechanics, 1,171
Dobler W., Haugen N. E., Yousef T. A., BrandenburgA., 2003, Phys. Rev. E, 68, 026304
Federrath C., 2013, MNRAS, 436, 1245
Federrath C., 2016, MNRAS, 457, 375
Federrath C., Klessen R. S., 2012, ApJ, 761
Federrath C., Klessen, Ralf S., Luigi L., Beattie J. R.2018, Nature Astronomy
Ferriere K. M., 2001, Reviews of Modern Physics, 73,1031
Hacar A., Tafalla M., Forbrich J., Alves J., MeingastS., Grossschedl J., Teixeira P. S., 2018, A&A, 610
Kainulainen J., Beuther H., Banerjee R., Federrath C.,Henning T., 2009, A&A, 2
Kolmogorov A. N., 1941, Doklady Akademii Nauk Sssr,30, 301
Konstandin L., Federrath C., Klessen R. S., SchmidtW., 2012, Journal of Fluid Mechanics, 692, 183
Kritsuk A. G., Norman M. L., Padoan P., Wagner R.,2007, ApJ, 665, 416
Mac Low M. M., Klessen R. S., 2004, Reviews ofModern Physics, 76, 125
Mocz P., Burkhart B., 2018, MNRAS, 480, 3916
Scalo J., 1998, in Gilmore G., Howell D., eds, Astronom-ical Society of the Pacific Conference Series Vol. 142,The Stellar Initial Mass Function (38th Herstmon-ceux Conference). p. 201 (arXiv:astro-ph/9712317)
WallpapersWide.com 2018, The Starry Night HD Wall-paper, http://wallpaperswide.com/the_starry_
night-wallpapers.html
Is The Starry Night Turbulent?